53 lines
3.2 KiB
Markdown
53 lines
3.2 KiB
Markdown
# RGB mask example: start in main.cpp
|
||
|
||
Both SDKs automatically include `examples/cpp/main.cpp`, `examples/c/main.c`,
|
||
public headers, matching libraries, docs and prebuilt C/C++ examples. For Seg,
|
||
the examples ALWAYS compose RGB instance IDs in memory after inference. No
|
||
`--rgb-mask` switch or BMP output is required; the old switch remains accepted.
|
||
|
||
The C++ main file directly demonstrates all 16 public APIs: lifecycle, model
|
||
info, image inference, count/performance/common results, bounding/rotated boxes,
|
||
Seg geometry/binary masks, pose keypoints and RGB composition. Task-specific
|
||
getters run only for the corresponding model. `write_result_json` is solely an
|
||
optional file-format illustration; the main's mask flow does not depend on it.
|
||
|
||
Read the numbered steps in main.cpp: get each borrowed 0/255 mask and geometry;
|
||
pass ordered mask views to yolo_masks_compose_rgb_ids; own the resulting RGB888
|
||
buffer; read a pixel at y*stride+x*3; remove zero and duplicate channel values
|
||
to obtain the one/two/three instance IDs. IDs are result index+1 (1..255), not
|
||
class IDs or cross-frame tracking IDs. RGB_PIXEL prints a real foreground pixel
|
||
and decoded IDs; ID_PIXELS prints membership counts. The C main shows the same
|
||
mask/RGB flow with explicit malloc/free.
|
||
|
||
Background=(0,0,0), single=(id,id,id), two=(id1,id2,id1),
|
||
three=(id1,id2,id3), ascending IDs. More than three overlaps returns an error.
|
||
Holes/disconnected regions stay intact. No dilation is performed by the SDK.
|
||
Do your processing at the marked location before RGB storage is freed; keep
|
||
the result set alive while using borrowed binary masks. RGB data requires no
|
||
file saving. A retained common/rgb_mask.h is an optional utility for older
|
||
integrations; the primary C/C++ examples no longer delegate mask handling to it.
|
||
|
||
Examples (use .exe on Windows):
|
||
```
|
||
yolo_cpp_example --config models/primary/model.yaml --image examples/assets/sample.bmp --json result.json --seg-geometry fit
|
||
```
|
||
Geometry modes: rect, circle, mean, fit. Seg RGB composition runs for each mode.
|
||
For other task types, RGB composition is not requested.
|
||
|
||
|
||
## 掩膜膨胀（2026-10-24）
|
||
|
||
C/C++ 示例支持 `--mask-dilation N`，N 为 0～100 的整数，默认 0。
|
||
Seg 的每个实例独立调用公共库 `yolo_mask_dilate`，再合成 RGB 编号图；保留重叠。
|
||
范围限制在模型 ROI（含调用方 ROI 覆盖）或整张原图内。JSON 的掩膜范围和
|
||
`maskDilationRadius` 对应处理后的数据；原始检测框和拟合几何保持原值。
|
||
公共库推理接口本身仍返回原始掩膜，RGB 合成函数本身不做膨胀。
|
||
|
||
调用流程：先用 NULL 输出、stride=0、capacity=0 查询返回尺寸与原点；
|
||
由调用方分配 width*height 字节，再调用一次得到二值掩膜。
|
||
保留结果集直到原始输入读取结束；膨胀输出由调用方持有并释放。
|
||
圆核满足 dx*dx+dy*dy <= N*N；N=0 保留原像素。输入输出不可共用缓冲区。
|
||
失败不改变输出及元数据，成功不改变行尾填充。可直接处理二值数据，无需保存图片。
|
||
RGB 最多保存三个重叠编号的限制保持不变；膨胀本身没有这个实例重叠限制。
|
||
可选膨胀耗时在推理计时之外；SDK/ARM64 交叉编译不能代替 NPU 实机验收。
|